NSA advisory urges silent downgrades for suspected distillers, clashing with Anthropic's June transparency promise
MysteriousAvocado580 · reddit · 2026-10-05
A detailed Reddit post examines the tension between Anthropic's June apology and the Sept 8 NSA/CISA/FBI advisory AA26-251A:
- June: after Fable 5's system card revealed flagged requests were silently answered by Opus 4.8, Anthropic promised Fortune that "any flagged requests will return a reason... You will see this every time it happens"
- The advisory on Chinese distillation campaigns explicitly recommends not informing suspected users of downgraded models, and varying changes to avoid triggering alerts (e.g., reduced reasoning depth)
The core problem is detection: listed indicators like 24/7 usage without idle periods, max-usage new subscriptions, and cache-optimization-tuned traffic are a fair description of any well-tuned scheduled agent fleet. The author, running a small multi-agent setup, was hit twice by reasoningextraction refusals on legitimate work (a text-leak bug discussion) — only caught because the refusal carried a label. A silent downgrade would have looked like "a bad day."
Contrasts: Anthropic's transparency page shows 11.4M banned accounts with appeal paths; OpenAI's Sept 30 shutdown of a Moonshot-linked reasoning-replay campaign used visible bans; Fable 5.1's anti-distillation tightening is visible too. But the author can find no statement since Sept 8 on whether "every time it happens" still holds.
More from Models
- AI Community Drama: Hermes Model Blasted as "Slop Just as Bad as OpenClaw" — alexandr_wang · 2026-10-05
- Reward hacking in the wild: models that reason about their grader narrow real-world judgment — thebasepoint · 2026-10-05
- Codex output on design work swings wildly between pro and intern quality — amitabhverma · 2026-10-05
- YC-backed inference provider accused of swapping models: its 'GLM-5.3-Flash' endpoint self-IDs as GLM-4.6 — NiceAd358 · 2026-10-05
- Multiple US open-weight frontier models dropping soon, says Bindu Reddy — bindureddy · 2026-10-05
- Opus 5.5 inside Hermes outshines Grok Bot and Dots in user hands-on test — EXM7777 · 2026-10-05